A diameter bound on the exponent of a primitive directed graph

A diameter bound on the exponent of a primitive directed graph

NORTH- HOLLAND A Diameter Bound on the Exponent of a Primitive Directed Graph Stewart W. Neufeld* The University Winnipeg, of Winnipeg MB R3B 2E9 ...

1MB Sizes 2 Downloads 44 Views

NORTH- HOLLAND

A Diameter Bound on the Exponent of a Primitive Directed Graph Stewart W. Neufeld* The University

Winnipeg,

of Winnipeg

MB R3B 2E9 Canada

Submitted by Richard A. Brualdi

ABSTRACT A directed graph G is primitive if there exists a positive integer /c such that for every pair of vertices u, w E G there is a walk from u to v of length k. The least such k is called the exponent of G. We define Gk to be the directed graph having the same vertex set as G and arcs (u,v) if and only if there is a walk in G of length k from vertex u to vertex ~1.A well-known upper bound for the exponent of a primitive directed graph G of order n is (n - 1)’ + 1, due to H. Wielandt in 1950. Our main result is the following refinement of the Wielandt bound: If G is a primitive directed graph with diameter d, then the exponent of G is at most d” + 1. We construct the primitive graphs for which equality is attained, and we generalize the bound to the class of irreducible matrices. In the course of proving the main result we find the following, which is interesting in its own right: If G is a primitive directed graph with diameter d, then the diameter of G” is at most d for all positive integers k.

1.

PRELIMINARIES

1.1.

Definitions and Notation

We generally follow the notation given in Brualdi definitions not contained in this section see (11. 1.1.1. primitive

entries.

Nonnegative

Matrices

A nonnegative

and Ryser

(entrywise)

[l]. For

matrix

A is

if there exists a positive integer k such that A” has all positive The minimum such k is called the exponent of A, which we denote

by exp(4. *This paper forms part of the author’s Ph.D. dissertation written at Queen’s 1Jniversity, Kingston under the supervision of D. A. Gregory and N. J. Pullman.

LINEAR ALGEBRA AND ITS APPLICATIONS 245:27-47 (1996) 0024-0255/96/$15,00 @ Elsevier Science Inc., 1996 SSDI 0024-3795(94)00203-P 655 Avenue of the Americas, New York, NY 10010

STEWART W. NEUFELD

28

It is evident that primitivity entries

does not depend on the size of the positive

in A, but only on their

positions.

loss of generality that all of our matrices J the all l’s matrix. DEFINITION 1.1. A (0,l) tation matrix P such that

we will suppose matrices.

All A12

= [

A22

A21

with no

We denote

A is reducible if there exists

matrix

PAPT

Hence, are (0,l)

by

a permu-

1

blocks All and A22 are square and Azr is a (nonvacuous)

where the diagonal block of zeros. A (0,l) matrices

matrix

which is not reducible

are included

which is irreducible

but not primitive

of imprimitivity k 2 2. We use the following result, lary 4.7 following

PROPOSITION is a permutation

Theorem

1.1. matrix

PAPT

=

matrices.

is said to be imprimitive

due to G. Frobenius

Primitive

A (0,l)

matrix

with index

[3], to establish

Corol-

4.1.

Let A be an imprimitive

(0,l)

matrix.

Then there

P such that 0

0

0

...

0

Al,

Al

0

0

..’

0

0

0

A2

0

‘..

0

‘.

:

.

.

0

0

...

A,c+~

0

0

0

0

“’

0

Ale-1

0

where k >_ 2 and the only l’s PAkPT

is said to be irreducible.

in the class of irreducible

occur in the blocks Al,

is the direct sum of k primitive

, Ak. Moreover,

matrices.

1.1.2. Graph Theory Terminology and Notation A directed graph G = (V, E) is a set of points, V, called vertices and a set of ordered pairs of vertices, E, called arcs. All of our directed graphs are finite, and we allow loops but no multiple arcs. The cardinality of V is called the order of G. Let U,V be two vertices of G. If (‘IL,U) is an arc of G, we say there is an arc from u to U. The number of arcs entering (leaving) a vertex w is called the indegree (outdegree) of U.

EXPONENT

OF A PRIMITIVE

A walls W necessarily

from

distinct

DIRECTED

GRAPH

u to u (or an u 4 vertices

V(W)

=

29

v walk)

is a sequence

( U, al, a~,

of not,

: ,u) and a set of arcs

a~), , (a,_ 1, a%), , (a3, I!). Since we do not, al( al, E(W) = (‘&al), low multiple arcs, specifying the vertex sequence of a walk uniquely dcscribes that walk. A closed walk is a walk where u = I’. A path P is a

walk where all the vertices all the vertices tion

‘II 5

except

‘0 to mean

are distinct.

there

IV = A + L3 is obtained W’ = kB indicates

are distinct.

We use t.he nota-

is a walk from ‘u to 1’of length

X:. The

walk

the final vertex of A with the initial

by identifying

of B. If Ic is a positive

vertex

A cycle C is a closed walk where

the first and last

and B is a closed walk, the notation

integer

the walk W obtained

by traversing

the closed walk E k

times. length of a walk W, denoted

The

IW], is the number

of arcs in kV. The

distance from u to V, denoted d(u, v), is the length of a minimum 11 + 19 path. We define d(u, U) = 0. If there is no IL + 71 path, then d(~: U) is not defined. If u and v are vertices

on a walk W, then W(U, U) denotes = ]E’(u,

U. u

u and u as members

on

a

walk W, we regard

even though,

for convenience,

we have dropped

A graph G is said to be strongly path

connected

from ‘U t,o v for all U, u E G. The

graph G is max{d(zl, diameter

Clearly,

(or strong)

diameter

the diameter

The girth of G is the length of a shortest

V(W).

of the sequence

the subscripts. if there exists

of a strongly

U) /U, u E G}. If G is not strongly

is not defined.

the portion

u)]. H ere, when we speak of vertices

of W from u to V, and l(u,~)

a

connected

connected,

then the

of a graph is at most II - 1.

cycle in G. If G has

no

the

cycks,

girt,h is not defined. Let L(G) denote the set of distinct cycle lengths L in G such that d + 1 5 L < 2d + 1. where d is the diameter of G. Let X(G) = IL(G)]. Similarly, define L(u)

to be the set of lengths

L of cycles in G wh,ich

co71hzi7~

u

and

such that d + 1 L: L 5 2d + 1. Let X(U) = ]L(u)]. We define G” to be the directed

graph with the same vertex

set as G

and arc

(u, V) if and only if there is a u 3 ‘~1walk. The exponent of is the least integer k such that for each pair r1.z: of G: denoted exp(G),

vertices

of G. there

integer

k such that

exp(G) = ma,x,,{exp(G; If G is the directed

is a u 3

v walk.

for each vertex

Define

exp(G;

%r in G there

u)}. graph whose adjacency

U) to be the least,

is a II 3

matrix

U. Clearly:

is the zero-nonzero

pattern of a nonnegative matrix A, then we say G is the directed graph associated with A. If G has diameter d, then we also say its adjacency matrix has diameter d. Tile following are well known; for example, Ryser

[11.

see R. A. Brualdi

and H. .J.

30

STEWARTW.NEUFELD

PROPOSITION 1.2. The (i, j) entry of A” is positive is a walk of length k from i to j in G = G(A).

if and only if there

Thus, A is primitive with exponent k if and only if G = G(A) is primitive k. with exponent In many of our results we will implicitly make use of the following characterization of primitivity: PROPOSITION 1.3. A directed graph G is primitive if and only if it is strong and it contains two closed walks of coprime length. 1.2.

Background

Wielandt [17] in 1950 provided (but did not prove) an upper bound on the exponent of a primitive matrix A in terms of n, the order of A. Proofs were later supplied by Heap and Lynn [6], Rosenblatt [12], Holladay and Varga [7], Dulmage and Mendelsohn [2], Pullman [ll], Lendaris [8], Schwarz [13], and Lewin [9]. PROPOSITION 1.4. The exponent of a (0,l) primitive matrix of order n is at most (n - 1)’ + 1. Moreover, the unique (up to simultaneous permutations of the rows and columns) matrix W’, for which exp(W,) = (n- 1)2+1 is given by

Wl

=

w, =

Our main

result

THEOREM

[ll,

...

and

.. .

.. .

oo...

... 0

10

for n 2 3.

1

0

...

0

0

1

1

0

...

0

0

01

is a refinement

of the Wielandt

bound:

4.1. If G is a primitive directed graph with diameter d, then exp(G)

I d2 + 1.

We will refer to this bound as the diameter bound. In the course proving the diameter bound we also obtain the following result:

of

31

EXPONENT OF A PRIMITIVE DIRECTED GRAPH

THEOREM 2.1. If G is a primitive directed graph with diameter d, then the diameter of Gk is at most d for all positive integers k. Hartwig and Neumann [5] mention that the diameter bound was conjectured by Hartwig in an unpublished working paper. In [5] they examine the following conjecture: if m = m(A) is the degree of the minimum polynomial of a primitive matrix A, then the exponent of A is at most (m - 1)’ + 1. They note that if d is the diameter of G = G(A), then d 5 m - 1. Thus. the diameter bound implies this conjectured bound. Gregory, Kirkland, and Pullman [4] prove that if A is primitive, then the exponent of A is at most (b - 1)’ + 2, where b is the Boolean ranlc of A, that is, the smallest integer b such that A is the Boolean product of n x b and b x n Boolean matrices. They show if d is the diameter of G = G(A), then d 5 b. Thus, the diameter bound is finer than their bound when d < b. The proof of the diameter bound is followed by an easy generalization to the class of irreducible matrices. Also, we obtain a construction of those graphs which attain the upper bound of d2 + 1. Our results are presented in the language of directed graphs. 1.3.

The Frobenius-Schur

Index

prime p0Sitk.W Let al < a2 < .‘. < ak be relatively gcd(ar, , ak) = 11. It is known that if N is a sufficiently then for all n > N the equation

integerS

[i.e.,

large integer,

blal + bzaa + . . . + bkak = n

has a solution in nonnegative integers bl , ba, . . , bk. The least such number N is called the Frobenius-Schur index and is denoted $(ar , a2, . . , ak). For k = 2 it is well known that 4( ar,a2) = (al - l)(az - 1). For Ic > 3 only some estimates are known. We cite two upper bounds on the Frobenius-Schur index due to Y. Vitek ([5, Theorem 41 and [16, Theorem 21, respectively). . < ak be relatively prime PROPOSITION 1.5 (Vitek). Let al < aa < positive integers. Let i be the Jirst index such that a, # Xal for integral X. If there is an aj # pal + uai for all nonnegative integers p, u, then

d(al,. . ,ak) 5 la1/21(ak - 2). Let al < a2 < . < ak be relatively prime having different residues module al. If, for every divisor

PROPOSITION 1.6 (Vitek). positive integers

STEWARTW.NEUFELD

32

r of al such that r < k - 1 and r does not divide k - 1, the number residues modulo al/r in {al, . . , ak} is not 1 + \(k - l)/rJ, then

of

For later use, we note the following immediate corollaries to the previous two propositions. COROLLARY 1.7. Let al < ... < ak be relatively with ak < 2al. Then

4(al,. . . , ak) 5

lal/21 (ak -

prime positive

integers

2>.

COROLLARY 1.8. Let al < a2 < a3 < a4 be relatively prime positive integers with a4 < 2al. If al is odd or if al is even and {al, a2, as, ad} does not have exactly two distinct residues modulo a1/2, then

4(al, a2, a3, a4) I

I 1 F

(a4 - 3).

Throughout the remaining sections we assume, unless otherwise specified, that G is a primitive directed graph, d is the diameter of G, and s is the girth of G. 2.

PROOF OF THEOREM

2.1

A proof of this theorem is contained in Neufeld [lo]. The theorem has been independently proved by Shen [14, Theorem 3.11 and it is his elegant formulation that we present below. THEOREM 2.1. Let G be a primitive directed graph with diameter Then the diameter of Gk is at most d for all positive integers k. Proof.

d.

Let u,v E G, u # v. By the primitivity of G we have a walk

u 3 Y for some r. Let rg = min{r: u 2 v}. We will show rg 5 d. Suppose to the contrary rg > d + 1. There are k + 1 vertices u = uo, Ul,. .., uk = v in G such that 7% RI T0 RI 771 u = u,, t ur 4 ‘9 -+ .” 4 uk-_l 4 ‘1Lk= v. Let 5, = d(ui_r,ui), 1 5 i 5 k. Then 0 < zi 5 d < r-0. Let yi = rg z, (mod k), 0 < yi 5 k - 1. Consider the set {cf=, yz: 1 = 1,2,.. ,k}. Then one of the following must hold:

EXPONENT

OF A PRIMITIVE

CASE

1.

There

exists some lo such that

CASE

2.

There

exists II < 12 such that

CtZ=ll+,

DIRECTED

GRAPH

3s

cfs, cfl,

yz E 0 (mod Ic). y, = C:=,

yz (mod k). i.e..

y2 = 0 (mod k).

Therefore, with no loss of generality for two integers 1 2 m. The walk

we suppose

that cl”=, yz = 0 (mod k)

(which has length
PROOF OF THEOREM AT MOST d

4.1.-THE

in G” we have ‘u 2 ‘1:;i.e., the W

CASE OF GIRTH

In this section we prove Theorem 4.1 in the special case where the girth of G is at most d. Surprisingly, the more restricted case where the girth of G is d + 1 appears to be more difficult. We deal with that lengthy case in the next section. LEMMA

girth

3.1.

Let G be a primitive

directed

graph

rwith. diameter

d and

s.

(a) If s < d, th en exp(G) (b) 1f s = d, th en exp(G) is a vertex

5 d2. 5 d2 + 1. &loreover

u E G such

cxp(G; u) = d2 + 1, then there cycle

of length

equality

that u is on no cycle is an arc (ILL,7:)

h.olds on.ly if th,ere of length

of

d. Also,

ij

G wh,ere 7~ is o’n a

d.

Proof. We first suppose d = 1. Then G2 has a loop at every vertex and has diameter equal to 1, and hence A” = ,I, where A is the adjacency matrix of G. Therefore, exp(G) < 2 = d2 + 1. We now suppose d >_ 2. Let u, ‘I’E G. Let C be a cycle of length s.

(a): Suppose s 5 d - 1. In G there is a u 3 w walk for some vert,ex w E C. By Theorem 2.1, GS has diameter at most d. Thus, a shortest w --t 11path in GS has length at most d, and since w has a loop in G”, there is a w 2 v walk. Hence, in G there is a w 2 1’walk and so a u ‘fsd v walk where d + sd 5 d + (d - 1)d = d2. Since TLand v are arbitrary vert,ices of G, we have exp(G) 5 d2.

STEWART W. NEUFELD

34

(b): Suppose s = d. We partition the vertex set V into sets S and T where S contains all vertices on cycles of length d and T contains all vertices on cycles of length d + 1 but not d. We show there is a u dl+! v walk in G. We have two cases. CASE

1.

Suppose u E S. Now in Gd, vertex u has a loop, and by Theo-

rem 2.1, the diameter of Gd is at most d. Thus, in Gd there is a u 3 v walk, and hence in G there is a u s exp(G; u) 5 d2. CASE

2.

v walk. Therefore,

since v E G is arbitrary,

Suppose 21E T.

(i) Suppose (u, w) is an arc of G and w E S. Then in Gd vertex w has a

dZ

loop and from case 1 there is a w + v walk. Therefore,

there is a

‘u.d% Y walk for each vertex w E G, and hence exp(G; u) < d2 + 1. (ii) Suppose there is no arc from u to any vertex in S. Let C be a cycle of length d + 1 containing u. Choose w E S such that there is a w 3 v walk. Let d(+ w) = 2, 2 5 x 5 d. Then there is a u 4 w walk where L = (d - x)lCI + x = d(d + 1 - x) 5 d(d - l), We note that L c 0 (mod d). Therefore,

since

in Gd there is a u 5

x 2 2. w walk

where k 5 d - 1, and using the loop at w in Gd, we obtain a u 5

w

walk. Hence, in G there is a u d(d-l) w walk and so a u 5 v walk where k = d + d(d - 1) = d2. Therefore, since u E G is arbitrary, exp(G; u) 5 d2. Since u E G is arbitrary, we conclude exp(G) 5 d2 unless u is on no cycle of length d and there is an arc (u, V) of G such that w is on a cycle of ?? length d, in which case exp(G) 5 d2 + 1.

4.

PROOF OF THEOREM EQUAL TO d + 1

4.1-THE

CASE OF GIRTH

The primary difficulty in proving the main result (Theorem 4.1) lies in the consideration of those primitive directed graphs G with diameter d and girth d + 1. In this section we will prove exp(G) 5 d2 for this remaining case. Throughout this section we will be dealing exclusively, therefore, with primitive directed graphs of diameter d and girth s = d + 1. We first observe that if G is a primitive directed graph, then there are closed walks in G of all lengths greater than or equal to exp(G). In particular, if m 2 2 is a positive integer, then there must be a closed walk in G

EXPONENT OF A PRIMITIVE DIRECTED GRAPH

35

whose length is not divisible by m. Moreover, every vertex of G must be contained in some such walk. In Lemma 4.1 we find, given a positive integer m 2 2 and a vertex u E G, an upper bound on the length of a shortest closed walk containing IL which is not divisible by m. LEMMA 4.1.

Let G be a primitive

u E G, and let m > 2 be a positive

directed graph with diameter

integer.

Then u is contained

d. Let

in. a closed

walk W where JWI 5 2d + 1 and IWI is not divisible by m. If the girth of G is d + 1, then W is a cycle. Let W be a shortest walk among all the closed walks in G which Proof. contain u and have lengths not divisible by m. Suppose, contrary to the statement of the lemma, that (WI > 2d + 1. Let W = W(u, v) + W(v, u), and choose v E W so that /W(u, v) / and (W(v, u) 1 both exceed d. Let Pi and P2 be shortest paths from u to w and from w to u respectively. Then W(% v) + p2, Pl + WV, u), and PI + Pz are all closed walks and are all shorter than W. Thus, IW(u,w)l

+ /P2[ E IPd + jW(u,u)(

= IPd + /Pzl = 0 (mod m)

But this implies IW(u, v)I + (W(v, u)I s 0 (mod m), a contradiction. If the girth of G is d + 1, then W is a cycle; for if it were not, then would contain a cycle of length < d + 1.

G ??

Lemma 4.1 implies each vertex u E G is on a cycle of length L where d + 2 5 L < 2d + 1 (choosing m = d + 1). We note that, since the girth of G is d + 1, every arc of G is contained in a cycle of length d + 1. DEFINITION 4.1. Let u E G. Let C be a cycle of shortest length among U. Similarly, define C’ all cycles in G of length at least d + 2 containing to be a cycle of shortest length among all cycles which contain u and have length greater than ICI. In the next two lemmas we obtain further and lengths of cycles contained in G.

information

about

the number

LEMMA 4.2. Suppose G is a primitive directed graph with diameter d 2 2 and girth d + 1. Let u E G. Let C be as in Definition 4.1. Then (Cl 6 (5d +4)/3. Proof.

at least d+2

By Lemma 4.1 every vertex in G is contained in a cycle of length and at most 2d+l. Let ICI = d+h, 2 5 h 5 d+l. Let a, b E C

STEWART W. NEUFELD

36

be vertices such that l(a, U) = h - 1 and 1(~, b) = h - 1. Let P, Q, and R be shortest u + a, a ---f b, and b + u paths. Now the cycle P + C(a,u) containing u is of length IPl + h - 1 where d + 1 5 IP( + h - 1 < d + h, and hence by the minimality of C we have )PI i h - 1 = d i- 1 and so (PI = d + 2 - h. Similarly, IR\ = d + 2 - h. We have l(b, u) = /Cl - l(u, u) l(u, b) = d + h - (h - 1) - (h - 1) = d + 2 - h. We note I&I 2 h - 1 or else Q + C(b, u) would be a cycle of length at most d, a contradiction. Therefore, the closed walk P + Q + R has length (Pl+~Q~+IRl>2(d+2-h)+h-1=2d+3-h>d+2,sinceh
and so ICI = d+h

5 (5d+4)/3,

which completes ??

LEMMA 4.3. Suppose G is a primitive directed graph with diameter 2 and girth d + 1. Then X(u) > 3 for all u E G.

d 2

Proof. By Lemma 4.1, X(u) > 2. Suppose, contrary to the lemma, that X(U) = 2. Let the cycles in L(u) have lengths d + 1 and d + h where 2 5 h 5 d + 1. Let (‘w,u) be an arc of G.

CASE 1. Suppose ‘w has outdegree at least 2. Let (w,v) be an arc of G, v # u. Let P, P’ be shortest 8 4 u and u --f II paths. Let R and S be shortest w + w and u + w paths. Then lRI = ISI = d, since the girth of G isd+l. The cycles P + S + (w, v) and P’ + S + (w, u) both have length d + h (since their lengths are at least d + 2 and at most 2d + l), which implies IPI = IP’l = h - 1. B u t now the cycle P + P’ has length 2h - 2, and since X(u)=2,wehave2h-2=d+lor2h-2=d+h.If2h-2=d+l,then h = (d + 3)/2, which implies gcd(d + 1, ICI) > 1, contradicting X(u) = 2 by Lemma 4.1. If 2h - 2 = d + h, then h = d + 2, contradicting the condition on h. Therefore, X(U) > 3. CASE 2. Suppose w has outdegree 1. Among all vertices in G which have outdegree at least 2, let z be one for which d(z, U) is minimal. Let (z, t) be an arc on a shortest z + u path. Then from case 1, A(t) 2 3. Since every cycle containing t also contains u, we have X(U) 2 3. ?? LEMMA 4.4. Let G be a primitive and girth d + 1. Let u E G.

directed graph with diameter

d > 2

EXPONENT

OF A PRIMITIVE

DIRECTED

(a) Let C be a cycle containing

GRAPH

U. JffiCl < 2d-

then exp(G; U) 5 8. (11) Let C and C’ be as in Dejinition tkn. cxp(G; Proof.

(a):

the vertex

:i 7

4.1.

Jf

d + 1 is odd and /C’/ < 2d.

Suppose

gcd(d + 1, iCl)

II E G” is contained

= :I’. 2 5

d. Then

exp(G”;

maximum and :I: =

on the interval (d + 1)/2.

2 5

Then

2.1, the diaIneter

of G,’

then

exp(C;

Th e f uric.t’Ion f(z)

f(s).

.f(:c)

II) 5.

attains

<: (d + 1)/2 at the endpoints

z

If :I: = 2, then

d > 2. If :I’ = (d + 1)/2,

f(x) = cl”~ d/2 + G < d” 4 1 = d2/2 + 2d -

5 < #

its

.I’ -: 2 foI

+ 1. Hc~ccs

‘u) 5 d2.

(11): Suppose we obtain

d + 1 is odd and lC”I < 2d. Then,

LEMMA 4.5.

Let G be a primitil~e

Let

if d >

JC/ =

5, then

directed graph. with diameter 4.1.

1.7. ??

rl 1 2

nrrn s~rppo.~r

‘: d’.

d + h, 2
Corollaq-

~ 2) + d = r?.

G. Let C be a,s in Definition

/Cl < 2d - 1. Then exp(G;u) PmoJ

applying

t:xp(G; U) 5 4(d -t 1, ICI, IC’l) + d < (d/2)(2d

an.d girth d + 1. Let u t

that

(d + 1)/2.

(d + 1)/z and ;C~/.I..

u) < q6((d + 1)/n,, ICl/n:) + rl = [(d + l)/.c. ~

(d + 1)(2d - 1)/n: - 3d + s(d + 1) =

exp(G:

.I: 5

in cycles of lengths

- 1) + d 5 (d + 1)(2d - l),‘.z2 - 3d/r + d + 1. Thus.

l](iCl/.I:

ICI) > 1.

U) < 8.

and gcd((d + 1)/z: ICI/z) = 1. Also from Theorem is at most

1 and gcd(d+l.

condition

jC(

<

2d -

by Lemma

1 is satisfied.

3.2

54:~~mwy

suppose gcd(d + 1, ICI) = 1, or else we are done by Lemma 4.4(a). By, Lemma 4.3 we have X(u) 2 3. We may also suppose X(71) = 3; othrmvisc, L(U) 2 {Cl,C2,C3,C4}, d+ 1 = ,C,/ < lC,i i IC:j < lC
4.1. Let ICY’1= d + k. 3 5 k < dt

1. Lrt ju’. II)

IX> an arc’ of C. CASE 1. Suppose w has outdegree at least 2. Let (w. (: # II. Let P, P’ be shortest v u and u -i II paths. shortest (1 + UJ and u --f w paths. is d + 1.

Then

7~)

be an arc of G. Let R and S 1~1

IRl = ~SI = d. since the girt,11 of G

Thecycle P+S+(w,v) has lengthd+l+jPi whercd+2 5 d+l+IP: < 2d i- 1. which implies IPI = h - 1 or IPl = k ~ 1, since X(M) = 3. If lP1 = h - 1, then IP’J = d + 2 - 1~. since the girth of G is d + 1 and lPI + lP’I

< ICI. If JPI = k -

1, then since

IPI + lP’1 = d + 1 or JPl + JP’I = IP’I =ct+/z-k+1.

IPl + 1P’I < IC’I and liencc~

K1, we 11iwc IP’l = d + 2 - k 01

38

STEWART W. NEUFELD

Choose 5 E C such that ]C(u,z)] = h - 1. Let T be a shortest z + u path. Then ITI = d+2-h, since thegirthof G is d+l and jT(+(C(u,x)1 < ICI. Let Q, Q’ be shortest Y -+ x and x --f v paths. (a) SupposeJP]=h-l.Then]P’]=d+2-handsothecycleP’+R+ (w, U) has length 2d+3-h. Since X(U) = 3 and 2 5 h 5 d-l, we have 2d+3-h=d+k,or2d+3-h=d+h.If2d+3-h=d+h,then h = (d + 3)/2, which implies gcd(d + 1, ICI) > 1, a contradiction, since we are assuming gcd(d + 1, ICI) = 1. If 2d + 3 - h = d + k, then h + k = d + 3. Since h < k, this implies h 5 (d + 2)/2. Then the cycle P + C(u, w) + ( w,v) has length d + 2h - 1, and since ]C] < d + 2h - 1 5 2d + 1, we have d + 2h - 1 = d + k. Then 2h - 1 = k or 3h - 1 = h + k = d + 3 and so h = (d + 4)/3. But now d + h = 4(d + 1)/3 and hence gcd(d + 1, ]C() > 1, a contradiction. (b) Suppose ]P] = k - 1. Then ]P’] = d + 2 - k or (P’] = d + h + 1 - k. If d = 3 and ]C] = 5 and ]C’( = 6 = 3(d + 1)/2, then the cycle P+C(u, w) + (w, v) has length 7, which contradicts X(U) = 3. Hence, in the following discussion we suppose if ]C’( = 3(d + 1)/2 then d > 3. (i) Suppose (P’I = d + h + 1 - k. Then the cycle P’ + R + C(w, u) haslength2d+2+h-kwhered+2I2d+2+h-k<2d+l, which implies 2d + 2 + h - k = d + k or 2d + 2 + h - k = d + h. If 2d + 2 + h - k = d + h, then k = d + 2, contradicting the condition on k. If 2d + 2 + h - k = d + k, then 2k = d + 2 + h. Then d + h is even, which implies d + 1 is odd [otherwise gcd(d + 1, ICI) > 1, a contradiction]. Since d + 1 is odd, we have k = d + 1, or else we are done by Lemma 4.4(b). But by assumption ICI 5 2d - 1, and so d + 2 + h 5 2d, and hence k 5 d, a contradiction. (ii) Suppose IP’J = d+2-k. Then the cycle P’+R+(w,u) has length 2d+3-k, and since 3 5 k 5 dfl, we have d+2 <2d+3-k < 2d and hence 2d + 3 - k = d + k or 2d + 3 - k = d + h. If 2d + 3 - k = d + k, then k = (d + 3)/2 < 2d - 1, for d > 3 and gcd(d + 1, IC’l) > 1, and so we are done by Lemma 4.4(a). Suppose 2d + 3 - k = d + h. Then h + k = d + 3. The cycle Q + C(x, w) + (w, v) has length at least d + 2 and at most 2d + 1, and so we have ]&I = h - 1 or IQ] = k - 1, since X(U) = 3. If IQ] = h - 1, then the cycle Q +C(x,u) +P’ has length 2d + 2 + h - k and we are done by the argument in (b)(i). If(Q]=k-l,then]Q’(=d+2-kor]Q’(=d+h+l-k,since X(U) = 3 and k - I+ IQ’1 < IC’(.

EXPONENT OF A PRIMITIVE DIRECTED GRAPH

39

If IQ’/ = d + 2 - k, then the cycle Q’ + R + C(w, x) has length (d+2-k)+d+h=2d+2-th-kandwearedonebytheargument in (b)(i). If IQ’/ = dfh-tl-k, then the cycle Q’+P+C(u, X) has length (d + h + 1 - k) + (k - 1) + (h - 1) = d + 2h - 1: and since h + k = d + 3, we are done by the argument in (a). Hence,

in case 1, we have exp(G; u) 5 d2.

CASE 2. Suppose w has outdegree 1. Among all the vertices in C with outdegree at least 2, let .z be one for which (C(z, u)] is a minimum. Let (2, t) be an arc of C. Then from case 1, exp(G; t) I: d2. Since every cycle ?? containing t also contains u, we have exp(G; u) 2 d2. We now consider

the small diameter

cases, that

is, d = 2,3,4.

LEMMA 4.6. Suppose G is a primitive directed graph with diameter d = 2,3, or 4 and girth d + 1. Let u E G. Then exp(G; u) 5 d2. Proof. Let C and C’ be as in Definition 4.1. By Lemma 4.3 we have X(u) >_ 3, and we may suppose X(u) = 3, or else we are done by Corollary 1.8. If JC( 5 2d - 1, then we are done by Lemma 4.5. Therefore, we suppose ICI = 2d and IC’I = 2d + 1. (a) Suppose d = 2. Let u E G. Then L(u) = {3,4,5}, since X(u) = 3. We will show there is a walk of length 4 = d2 from u to every vertex of G. Clearly, there is a cycle of length 4 containing u, and there is a u 5 ZI walk for all vertices

u with d(u, u) = 1. It remains

to show

that there is a u 3 w walk where d(u, w) = 2. Let C : u, a, b, c, d, u be a cycle of length 5 containing u. Let Qi, Q2, Qs be shortest u ----)d, d -+ c, c -+ b paths. Then l&i/ = IQ21 = l&s1 = 2, since the girth of G is 3. Let PI, P2, P3, P4 be shortest u -+ c, c -+ a, a --+ d, and d --f b paths. We note that w # c; otherwise the path Qi + Q2 has length 4. Also, d(c, w) = 2, or else the path C(u, c) -t- (c, w) has length 4. Then [PiI = 1, or else the path consider two cases:

pi + c 3

w has length

4. We now

CASE 1. Suppose b # w. Then d(b, w) = 1, or else the path C(u, b) +b -% w has length 4. But now the path 4 + Qa + (b, w) has length 4. CASE 2. Suppose b = w. Then lP41 = 1, or else the path Qi + P4 has length 4. Also, [Pzl = 1, or else the path PI + P2 + (a, b) has length 4. As well, 1P3 1 = 1, or else the path (u, a) + P3 -t P4 has length 4. But now the path PI + P2 + P3 + P4 has length 4.

STEWART W. NEUFELD

40

(b) Suppose d = 3. We have L(u) = {4,6,7}. We note that if u is also contained in a closed walk of length 9, then exp(G; U) < 4(4,6,7,9)+ d = 9 = d2. Choose vertices a, b,c E C’ such that \C’(a, u)I = \C’(b,a)l = IC’(u,c)( = 2 and IC’(c,b)l = 1. Let P~,P~,P~,P~ be shortest u -+ a, a + b, b -+ c, and c ---) u paths. Then IPrl = 2, or else u is contained in a cycle of length 5, which contradicts the minimality of C. Similarly, IPd( = 2. Also, lP31 = 3, since the girth of Gisd+l=4.IfIPzJ=2,thenthecycleP~+Pz+P3+Pqhaslength 9 and we are done. If IPzl = 3, then the closed walk PI +Pz +C’(b, u) has length 9 and again we are done. (c) Suppose d = 4. We have L(u) = {5,8,9}. Choose vertices a, b E C such that IC(a,u)( = (C(u, b)l = 3. Then IC(b,a)I = 2. Let P, Q, R be shortest u --) a, a --f b, and b --f u paths. Then IPI = 2, or else u is contained in a cycle of length 6 or 7, contradicting the minimality of C. Similarly, IRI = 2. We note that I&I > 3, or else the cycle Q + C(b,a) has length <5 = d + 1. If I&I = 3, then the cycle P + Q + R has length 7, contradicting the minimality of C. Suppose I&I = 4. Then the cycle Q + C(a, b) has length 6. Let u E G. Then the u 4 v walk P + a --f v intersects cycles of lengths 5, 6, 7, and 9 and, since v E G is arbitrary, we have exp(G; U) < 4(5,6,8,9) + IPI + d = 14 < d2. ?? THEOREM 4.1.

Suppose G is a primitive

directed graph with diameter

d. Then

exp(G)

< d2 + 1.

Proof. If the girth of G is at most d, then we are done by Lemma 2.6. If the girth of G is d + 1, then by Lemmas 4.4, 4.5, and 4.6 we have exp(G; u) < d2. Since u is an arbitrary vertex of G, we have exp(G) = ?? maxexp(G; u) < d2. We can generalize Theorem 4.1 to the class of irreducible matrices. Define the index of convergence, c(A), of an irreducible (0,l) matrix A to be the smallest integer c 2 1 such that for all j 2 c, Aj = Aj+p for some p 2 1 (where all arithmetic is Boolean). COROLLARY 4.7. Let A be an irreducible matrix with index itivity k and diameter d. Then the index of convergence c(A) d2/k + k. Proof.

of G(A)

Since A is irreducible with index of imprimitivity can be partitioned into k nonempty sets VI, V2,.

of imprimis at most

k, the vertices . , Vk where all

EXPONEIZT

OF A PRIMITIVE

in V, enter

the arcs originating square

diagonal

DIRECTED

blocks

Let m = maxexp(B,),

B,.

Vi+,

GRAPH

II

(k + 1 = 1). Now A” consists

Also, by Proposition

1.1 each B,

of k

is primitive:.

1 < i < k. Then

(A”)”

1,

0

0

0

.J,

0

0

0

J3

0

0

0

=

where each ,I, is a square

matrix

.”

0 0

“.

0

.

.Jk

of 1’s. Thus,

c(A) 5 km.

Since t’he diameter of G(A) is d and every path between two vertices iu V,: 1 5 i < k, has length divisible by k, the diameter of B,. 1 5 i 5 k, is jd/kJ 5 d/k. By Theorem

at, most c(A)

if A is primitive,

We note that

then gives the result of Theorem From equality

5 (d/k)2 + 1. Therefore.

the lemmas

in Theorem

in this section

we construct

CONSTRUCTION

In

then k = 1 and c(A)

the

= exp(A).

which

4.1. and from Lemma

3.1, we note that

4.1 only if G has girth 5 d. In

the family of graphs for which equality

holds

4.1.

BOUND

exp(G)

??

can occur in the bound of Theorem

the next. section

5.

4.1, exp(B,)

< km 5 k[(d/k)’ + l] = d2/k + k.

OF GRAPHS

ATTAINIKG

THE

UPPER

d” + 1

following

we construct

= d2 + 1. We restrict

For if d = 1 and exp(G)

those

primitive

our attention

graphs

to the cases

= d2 + 1 = 2: then

it is easily

G for which where

d 2

seen that

2.

G is

a complete directed graph on n. vertices with at least one of the loops r+ moved. (If n = 2, then precisely one of the vertices has a loop; otherwiscl G is not, primitive.

If n > 3, loops are optional,

since even wit,h no loops G

is primitive.) For d 2 2, we describe a family, J-d. of graphs which has the propert!. that the exponent of each G E Fd attains the upper bound d2 + 1. TIM, family 3,l consists

(1) The

of the following directed

vertex set of G is Vo U VI U.

graphs

G.

U Vd. where the V, are pairwise

joint and nonempty and Vo consists of a single distinguished (2) The arc set of G is defined as follows:

dis-

vertex II.

(i) For 0 < i 5 d, ( u, w) is an arc of G for each 21 E V, ancl C&I 1~1 E V,+, , where addition is taken modulo d + 1.

STEWARTW.NEUFELD

42

(ii) The remaining arcs in G may be any set of arcs from vd to VI with the following properties: for each vertex 2, E vd, (v, ‘UI)is an arc for some ‘w E VI, and for each v E VI, (w, w) is an arc for SOme

W

E

vd.

We observe the following about graphs G in the family F,j. (1) Each G in Fd has diameter d. (2) The distinguished vertex u is on a cycle of length d + 1 but on no cycle of length d. (3) The length of each cycle in G is a nonnegative linear combination of d and d+l. We will show the graphs in this family are the only ones (up to isomorphism) which attain the upper bound d2 + 1. (see Figure 1.) In the lemmas to follow, we assume that G is a primitive directed graph with diameter d > 2 and exponent d2 + 1, and we find properties of G that allow us to conclude that G is in the family Fd. LEMMA 5.1. Suppose G is a primitive directed graph with diameter Suppose exp(G) = d2 + 1. Then the girth of G is d.

d.

Proof. We note that, by Lemma 3.1(a), if the girth of G is at most d - 1, then exp(G) 5 d2. If the girth of G is d + 1, then by Lemmas 2.16

FIG. 1.

Graphs with exponent d2 + 1.

EXPONENT

OF A PRIMITIVE

DIRECTED GRAPH

and 2.17, exp(G) 5 d2. Therefore, we conclude that if exp(G) then the girth of G is d.

43

= d2 + 1. ??

Note that since G has girth d, every closed walk in G of length less than 2d must be a cycle. Also, if a vertex u of G is on no cycle of length d, then every closed walk that contains u and has length 2d must be a, cycle. LEMMA 5.2. Suppose G is a primitive directed graph with diameter d. Let u be a vertex such that exp(G; u) = d2 + 1. Then u is not on an,y cycle of length, t, where d + 2 < t 2 2d.

Proof. We note by Lemma 3.1(b) that if exp(G;u) = d2 + 1, then u, is on a cycle of length d + 1 and not on any cycle of length d. Also, by Lemma 3.1(b), there is an arc (u, v) where ‘V is on a cycle of length d. CASE I. Suppose u is on a cycle of length t where d + 2 < t 5 2d - 1. Let w E G be any vertex, and let P be a shortest v -+ w path. Then the u -+ w walk consisting of (u,v) + P has length at most d + 1 and meets cycles of lengths d and d + 1. Therefore, by Corollary 1.7, since vertex U: is arbitrary,exp(G;u) 5 $(d,d+l,ICl)+d+l < (2d-3)d/2+d+l < d2+1. a contradiction. CASE II. Suppose u is on a cycle of length 2d. We show there is a u 5 ‘~1 walk for each w E G. We observe by the argument in Lemma S.l(b)(ii) a shortest v + w path has length d, since we are assuming exp(G; ,u) = d2 + 1. Suppose d = 2. Then u is contained in cycles of lengths 3 and 4 = 2d. We wish to show there is a walk of length 4 from u to each vertex of G. Then we would have exp(G; U) < 4 = d2, which contradicts the exponent assumption on u. SUBCASE 1.

We first observe there is a u 5 u cycle and a u 3 w walk for each 111 such that (u, w) is an arc of G. It remains to be shown there are u 3 III walks where d(u, w) = 2. Let UJ be a vertex at distance 2 from U. Let (u,z;u)) be the vert,ex sequence of a shortest u 4 w path. If x is on a cycle C of length d = 2. then the walk (u, zr) + C + (z, w) has length 4 and we are done. Therefore. we suppose II:is not on a cycle of length d. By Lemma 3.1(b), there is an arc (u, V) such that z, is on a cycle of length d = 2. In particular, v # z. Let, P, P’ be shortest v --t x and x -+ v paths. Then we may suppose 1PI = 1, or else the walk (u,v) + P + (x, w) would have length 4. Then IP’l = 2 since .r is not on a cycle of length 2. Let (v, u, U) be the vertex sequence of

STEWART W. NEUFELD

44

a cycle C of length d = 2 containing u. Let Q, R be shortest y + w and z 4 y paths. If IQ1 = 0 (that is, w = y), then the walk (‘~1,X) + P’ + (v, w) has length 4. If I&I = 2, then the walk (qv) + (v, y) + Q has length 4. Suppose IQ1 = 1. If IR\ = 1, thenthewalk (u,u)+(~,~)+Rt& haslength 4, and if JR1 = 2, then the walk (u, z) + R + Q has length 4. Since there is a walk of length 4 from u to each vertex of G, we have exp(G; u) 5 4 = d2, a contradiction. SUBCASE 2. Suppose d > 3. We first show u is on a cycle of length 2d which also contains v. Let C be a cycle of length 2d containing u but not u. Choose w E C such that l(u, w) = d - 1. Let R,S be shortest w -+ u and u + w paths. Then (RI = 2, since by case I u is on no cycle of length twhered+2 4. Then ISI = d - 2; otherwise the cycle (u, w) + S + R would have length t where d+2 5 t 5 2d-1. But now the cycle (ZL,u)+S+C(w, u) has length 1 + (d - 2) + d + 1 = 2d. Suppose d = 3. If IS/ = 1, then the cycle (~L,‘u)+ S+C(w, u) has length 6 = 2d. If JSI = 2, then the cycle (u, V) + S + R has length 5 = 2d - 1 contradicting case I. If ISI = 3, then the cycle (21,V) + S + R has length 6 = 2d. SUBCASE 2(a).

Suppose w = u. Since ‘u, is on a closed walk of length

2d which also contains ZJ, all multiples of d are lengths of closed walks

containing u. Therefore, there is a u z UJwalk. SUBCASE 2(b). Suppose w # u. Let P,Q be shortest u ----fw and IJ----fw paths. Then I&I = d, or else we are done by the argument in Lemma 3.1 (b) (ii) Let C, C’ be cycles of lengths 2d and d + 1 containing ‘~1and V, and let C” be a cycle of length d containing v. If IPI = 1, then (d-l)C+( u, w) is a u % w walk. Suppose IP( 2 2. Let (u, x) be an arc on P. Let R, R’ be shortest x -+ u and IJ+ x paths. Let S,T be shortest IJ + u and x + u paths. Then /SI = d, since u is on no cycle of length d. Similarly, ITI = d. Now (RI = d - 1 or (RI = d, since by case I the cycle (u, z) + R+ S cannot have length t where d + 2 5 t 5 2d - 1. Similarly, IR’l = d - 1 or IR’I = d. If\RI=d-l,then(u,x)+R+Q.

is a u 2 w walk, and since it includes

‘u (u is on C”), there is a u c w walk. Hence, we suppose (R( = d. If lR’( = d, then the ZL---f w walk consisting of (d - JPI)C’ + (u,u)+ (IPI - 2)C” + R’ + P(x, w) has length (d - lPl)(d + 1) + 1 t- (IPI - 2)d + d + /P( - 1 = d2. Hence, (R’l = d - 1. But now the cycle R + R’ has length 2d - 1 and we are done by the argument in case I. ??

EXPONENT

1;Z’ewill exp(G;

OF A PRIMITIVE

DIRECTED

proceed to describe

now

U) = d2 + 1. Partition

where Cra = {u} partition

GRAPH

-1i

the arcs of G. Let u be a vertex such that

the vertex

set of G into sets Ua, Ui,

and IC E U, if and only if d(~,z)

there is an arc from u to each vertex

LE~~MA 5.3. Suppose

exp(G)

Suppose

G is a primitive

= i. By definition

. UC,. of the

in Ui

disrected ,qraph with dinmetcr

= d” + 1. Let u he a vertez

such

th,at exp(G;

d.

u) = (I” t 1

‘7%P77:

(a) Each (b)

vertex

There

in G is on. a qcle

is no arc from,

In particular,

there

U,

through

to Vi where

is n,o arc from

11 of lenqth

U, to Liz.

(c) .I’ t U, nnd y E Ul+lT 1 < i <: d (where d - 1) implies (d) Suppose

d + 1.

0 5 i 5 j < d n,rr.dj - i < rl -‘- 1 addition

is taken

rnotlulo

(x; y) is an arc oj’G.

x E Ud and :y E U1. Then.

(i) there is n w E UC{such that (w_ y) is nn (II% of G. nnd (ii) there Proojl shortest through Lemma

is a z E U1 such, that (x, z) ‘is an. arc of G.

(21): Suppose

.z’E U, for some i where

1 5 i 5 ct. Let, P, P’ br,

II ---) :c and FIZ--f u paths. Then jPl = i and P + P’ is a closed walk u of length at most. i + tl < 2d. T~IIS, P + P’ is a cycle. and I)\. 5.2. lPI + IF”1 = d + 1.

I\,‘c observe

that

(a) implies that there is an arc from each vertex

in F<,

to II. (1,): Suppose

to the contrary

that .c E U, and y E lJj and (y%:r:) is an arc

of G. By (a) there is a cycle C through through

u and x of length dt 1 and a cycle C’

u and y of length d + 1. Then C’( u, y) + (y. .x) + C(:c, U) is a closed

u~alktl~roughusatisfyingd+2~j+l+(d+1-i)=d+2+(~~-~)~2rl since 0 5 j - i < d - 1. This is impossible by Lemma

5.2.

(c): By (b), if (z,y) is not an arc of G> then the shortest x + y path is of the form P + (z, y) where z E U,, P is a shortest :I’+ 2 path, and (z, y) is an arc of G. But lPI = d, which implies d(x. g) = cl + 1. contradic:ting t,he diameter assumption on G. (d)(i): 1Ve have two cases: CASE 1. Suppose /Ui/ = 1. Then if there is no II! E ud such that (uj,yj is an arc of G, every cycle of G goes through u, and hence all cycles have lengt~h d + 1. Therefore,

G is not primitive,

a contradiction.

CASE 2. Suppose IUi / > 1. Let 11 be a second vertex in Ui. If there is no ‘W E Ud such that (w, y) is an arc of G, then every walk from v to y goes since G has through 7~ and hence has length at least d + 1, a comradiction, diameter

d.

STEWART W. NEUFELD

46

(d)(ii): By an argument similar to that z E U1 such that (z, z) is an arc of G.

in (i) we can show there

is a ??

THEOREM 5.1. Let 3d be the family of directed graphs described at the beginning of this section. Suppose G is a primitive directed graph with diameter d. Then exp(G) = d2 $1 if and only if G E 3;1. Proof. Suppose G E 3d. Then a consequence of the definition of 3d is that the lengths of all closed walks in G are nonnegative linear combinations of d and d+l. We note that since gcd(d, d+l) = 1, G is primitive. Moreover, there is a vertex ‘u.E G which is on cycles of length d + 1 but not of length d. The distinguished vertex ‘1~of G is contained only in closed walks of lengths sd + y(d + 1) where z 2 0 and y > 0. There is no u -+ u walk of length d2; for if xd + y(d + 1) = 0 (mod d), then y = 0 (mod d), which implies y 2 d and so xd + y(d + 1) 2 d2 + d. Therefore, exp(G) > d2 + 1. By Theorem 4.1, exp(G) 5 d2 + 1 and therefore exp(G) = d2 + 1.

Suppose exp(G) = d2 + 1. The descriptions of the arcs of G from Lemmas 5.1, 5.2, and 5.3 are precisely those which define graphs in the family 3d. Therefore, we conclude that if exp(G) = d2 + 1, then G belongs to the ?? family of graphs 3d. REFERENCES 1 2 3 4 5

6 7 8 9 10

R. A. Brualdi and H. J. Ryser, Combinatorial Matrix Theory, Encyclopedia Math. Appl., Cambridge U.P., 1991. A. L. Dulmage and N. S. Mendelsohn, Gaps in the exponent set of primitive matrices, Illinois J. Math. 8:642-656 (1964). G. Frobenius, fiber Matrizen aus nicht negativen Elementen, Sitzungsber. K. Preuss. Akad. Wiss. Berlin, 1912, pp. 456-477. D. A. Gregory, S. J. Kirkland, and N. J. Pullman, A bound on the exponent of a primitive matrix using Boolean rank, to appear. Robert E. Hartwig and Michael Neumann, Bounds on the exponent of primitivity which depend on the spectrum and the minimal polynomial, Linear Algebra Appl. 184:103-122 (1993). B. R. Heap and M. S. Lynn, The index of primitivity of a nonnegative matrix, Namer. Math. 6:120-141 (1964). John C. Holladay and Richard S. Varga, On powers of nonnegative matrices, Proc. Amer. Math. Sot. 9:631-634 (1958). G. Lendaris, Two Theorems Concerning Regular Markov Chains, Electronics Research Lab., Univ. of California, Berkeley, 1961. M. Lewin, On exponents of primitive matrices, Numer. Math. 18:154-161 (1971). S. Neufeld, A Diameter Bound on the Exponent of a Primitive Directed Graph, Ph.D. Dissertation, Queen’s University at Kingston, Canada, 1993.

EXPONENT

11 12

13 14 15 16 17

OF A PRIMITIVE

DIRECTED

GRAPH

47

N. J. Pullman, On the number of positive entries in the powers of a nonnegative matrix, Cunad. Math. Bzd1. 7:525-537 (1964). David Rosenblatt, On the graphs and asymptotic forms of finite Boolean relation matrices and stochastic matrices, Naval Res. Logist. Quart. 4:151157 (1957). Stephan Schwarz, On the semigroup of binary relations on a finite set, Unpublished manuscript, Bratislava, circa 1967. J. Shen, The proof of a conjecture about the exponent of primitive matrices, Linear Algebra Appl., to appear. Yehoshua Vitek, Bounds for a linear diophantine problem of Frobenius, II, Canad. J. Math. 28:1280-1288 (1976). Yehoshua Vitek, Bounds for a linear diophantine problem of Frobenius, J. London Math. Sot. 2(10):79-85 (1975). H. Wielandt, Unzerlegbare, nicht negative Matrizen, Math. 2. 52:642-645 (1950).

Received 21 December 1993; final manuscnpt

accepted 7 September 1994